+ case Intrinsic::x86_sse4a_insertqi: {
+ // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
+ // ones undef
+ // TODO: eventually we should lower this intrinsic to IR
+ if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
+ if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
+ if (CIWidth->equalsInt(64) && CIStart->isZero()) {
+ Value *Vec = II->getArgOperand(1);
+ Value *Undef = UndefValue::get(Vec->getType());
+ const uint32_t Mask[] = { 0, 2 };
+ return ReplaceInstUsesWith(
+ CI,
+ Builder->CreateShuffleVector(
+ Vec, Undef, ConstantDataVector::get(
+ II->getContext(), ArrayRef<uint32_t>(Mask))));
+
+ } else if (auto Source =
+ dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
+ if (Source->hasOneUse() &&
+ Source->getArgOperand(1) == II->getArgOperand(1)) {
+ // If the source of the insert has only one use and it's another
+ // insert (and they're both inserting from the same vector), try to
+ // bundle both together.
+ auto CISourceWidth =
+ dyn_cast<ConstantInt>(Source->getArgOperand(2));
+ auto CISourceStart =
+ dyn_cast<ConstantInt>(Source->getArgOperand(3));
+ if (CISourceStart && CISourceWidth) {
+ unsigned Start = CIStart->getZExtValue();
+ unsigned Width = CIWidth->getZExtValue();
+ unsigned End = Start + Width;
+ unsigned SourceStart = CISourceStart->getZExtValue();
+ unsigned SourceWidth = CISourceWidth->getZExtValue();
+ unsigned SourceEnd = SourceStart + SourceWidth;
+ unsigned NewStart, NewWidth;
+ bool ShouldReplace = false;
+ if (Start <= SourceStart && SourceStart <= End) {
+ NewStart = Start;
+ NewWidth = std::max(End, SourceEnd) - NewStart;
+ ShouldReplace = true;
+ } else if (SourceStart <= Start && Start <= SourceEnd) {
+ NewStart = SourceStart;
+ NewWidth = std::max(SourceEnd, End) - NewStart;
+ ShouldReplace = true;
+ }
+
+ if (ShouldReplace) {
+ Constant *ConstantWidth = ConstantInt::get(
+ II->getArgOperand(2)->getType(), NewWidth, false);
+ Constant *ConstantStart = ConstantInt::get(
+ II->getArgOperand(3)->getType(), NewStart, false);
+ Value *Args[4] = { Source->getArgOperand(0),
+ II->getArgOperand(1), ConstantWidth,
+ ConstantStart };
+ Module *M = CI.getParent()->getParent()->getParent();
+ Value *F =
+ Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
+ return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
+ }
+ }
+ }
+ }
+ }
+ }
+ break;
+ }
+
+ case Intrinsic::x86_avx_vpermilvar_ps:
+ case Intrinsic::x86_avx_vpermilvar_ps_256:
+ case Intrinsic::x86_avx_vpermilvar_pd:
+ case Intrinsic::x86_avx_vpermilvar_pd_256: {
+ // Convert vpermil* to shufflevector if the mask is constant.
+ Value *V = II->getArgOperand(1);
+ unsigned Size = cast<VectorType>(V->getType())->getNumElements();
+ assert(Size == 8 || Size == 4 || Size == 2);
+ uint32_t Indexes[8];
+ if (auto C = dyn_cast<ConstantDataVector>(V)) {
+ // The intrinsics only read one or two bits, clear the rest.
+ for (unsigned I = 0; I < Size; ++I) {
+ uint32_t Index = C->getElementAsInteger(I) & 0x3;
+ if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd ||
+ II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256)
+ Index >>= 1;
+ Indexes[I] = Index;
+ }
+ } else if (isa<ConstantAggregateZero>(V)) {
+ for (unsigned I = 0; I < Size; ++I)
+ Indexes[I] = 0;
+ } else {
+ break;
+ }
+ // The _256 variants are a bit trickier since the mask bits always index
+ // into the corresponding 128 half. In order to convert to a generic
+ // shuffle, we have to make that explicit.
+ if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 ||
+ II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) {
+ for (unsigned I = Size / 2; I < Size; ++I)
+ Indexes[I] += Size / 2;
+ }
+ auto NewC =
+ ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size));
+ auto V1 = II->getArgOperand(0);
+ auto V2 = UndefValue::get(V1->getType());
+ auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC);
+ return ReplaceInstUsesWith(CI, Shuffle);
+ }
+